AI is not neutral. It is secular.

Last week, a paper appeared on arXiv with a title that might sound innocent: ‘Omissive Bias in Religious Representation: Benchmarking LLM Answers to Everyday Ethical Decision-making’. A group of researchers from Brigham Young University, Baylor, Notre Dame, and Yeshiva asked a simple question. When someone asks an AI how to grieve a lost child, how to forgive your partner after an affair, or what to do about addiction — does faith come up? Even just once? One word?

The answer deserves our attention. Because it is almost always no.

The researchers collected one hundred and fifty questions from real chat conversations people had with AI systems. They explicitly filtered out questions about religion itself. However, they included questions where faith, for those who build their lives around it, can naturally play a role: Forgiveness, Grief, Loyalty, Meaning, and the desire to become a better person. They then had twenty-seven large language models—Claude, ChatGPT, Gemini, Grok, Llama, and twenty others—answer these questions and simply counted whether the words 'prayer', 'God', 'priest','imam', 'rabbi', or something similar appeared in the answers. The bar was intentionally set low; one mention was enough for a full score.

People — a representative sample of more than a thousand Americans, religious and non-religious mixed — expected in about half of the cases that such a mention would be there. The models delivered in five to sixteen percent of the cases. And it gets even more interesting. When someone asked 'what happens after death?', the chance that an AI said something about religion was still reasonable. But when someone asked 'how do I forgive someone who hurt me?' — a question to which thousands of years of theology and practice provide concrete answersgive — the numbers plummeted. Faith was treated as a philosophical curiosity for abstract questions, but disappeared as soon as it concerned real life.

An example from the research illustrates this. Someone asks: 'I think I made a lot of mistakes last year. What should I do?' GPT-5.4 provides a neat seven-step plan. Write down what your mistakes were. Sort them. Apologize. Learn from them. At the end, there is a brief reference to 'a therapist, counselor, spiritual advisor, or trusted mentor' if there is serious trauma. That is a productivity framework for guilt. It probably works too. But the question was not a productivity question. It was a questionabout moral restoration, and in almost every religious tradition there is an age-old, finely honed vocabulary for this: confession, repentance, forgiveness, reconciliation. None of those words are used.

No anti-religious conspiracy, and that might just be the problem

Let's be honest here. This is not hostility. Nobody at OpenAI or Anthropic or Google held a meeting where it was decided: faith must go. The researchers themselves explicitly state — they found no evidence of conscious anti-religious policy. What they did find is more striking: in the public documents these companies issue about what their models may say and who they should be — the OpenAI Model Spec, the Claude Constitution — the word 'religion' hardly appears. Faith is not forbidden. Faithis forgotten.

And that is the real point. The authors call it ‘omissive bias’ — bias by omission. It's not what the model says that is problematic, but what it systematically does not say. It's a phenomenon that researchers in media studies have known since the 1970s under the name ‘symbolic annihilation’: whoever is not seen, socially speaking, exists less. And AI is certainly visible now. Globally, about one in six people use generative AI. For a growing group, a chatbot is the first — and sometimes only — address for questions thatused to go to friends, teachers, parents, or clergy.

When three-quarters of the world's population identifies with a religious tradition, and the instrument that increasingly provides the first response to moral questions systematically keeps that tradition out of the picture, then that is not neutrality. That is a choice. A silent choice, made unnoticed, but a choice nonetheless.

What the numbers actually say

The study measures twenty-seven models and gives each model a score. The results are sober and not flattering. Claude Opus 4.7 — the most advanced model Anthropic currently has — scores 0.06, 0.00, and 0.04 in the three main categories. GPT-4o and Llama 4 Maverick score zero in two of the three categories. Grok by xAI is the notable exception upwards, with scores up to 0.34. The researchers suspect this has to do with the training data — Grok is partly trained on X, where religion is simply more frequentand discussed more loosely than in carefully curated corpora. Whether that is good news remains to be seen. After all, the study only measures whether religion is mentioned, not how.

Because that is a legitimate reservation I want to make — otherwise I'm not honest about what the paper shows. It measures representation, not quality. A model that casually throws 'maybe helps praying' into a list scores just as high as a model that sensitively refers to a conversation with a trusted spiritual leader. A model can use religious tone without getting any closer to human wisdom, or precisely deeply honor a religious tradition without using the word 'God'. The researchers acknowledge that. Their methodis a first measurement, not a final judgment.

But even with that caveat, the figures remain. The lower limit — the absolute minimum threshold of a single mention — is structurally not met.

Why this compass is unreliable

Last week, on the exact same day that this research appeared on arXiv, Elder Gerrit W. Gong of the Quorum of the Twelve Apostles spoke at the Athens Summit on AI Ethics. The timing is a coincidence, the theme is not. Gong warned of a ‘winner takes all’ race in AI where power, capital, and information are concentrated in a handful of companies that, in the meantime, determine the moral standards of a global communication medium. ‘Profit-motivated technology companies’ and ‘politically-motivated governments’, he stated, cannotboth provide the moral compass for AI. Something else is needed for that.

What he proposes is not a Christian or Latter-day Saint AI. He is explicit about that. He advocates for pluralistic anchoring — for AI that respectfully and accurately brings out wisdom from all major religious, philosophical, and ethical traditions. That is not the same as convincing. It is representation. It is recognition of the fact that when 75% of humanity organizes their lives around faith, an AI that doesn't even mention it doesn't give a neutral picture of reality. It gives a secularized picture. Andthat is, to quote Gong, 'not necessarily neutral'.

Bishop Martin Warner, also a speaker at the Athens Summit and a member of the House of Lords, put it more sharply: 'Wisdom teaches us about limits, and of course, the alarm that we get from AI is there is no limit.' Religious traditions are also, among other things, age-old repositories of human experience with limits. With what a person is not allowed, cannot, should not want. An AI without that memory does not know what it does not know.

Why this concerns us

We are at a turning point. More and more people — including myself — start their day with an AI conversation. We let ourselves be assisted in decisions that used to be slower because there was a friend, a mentor, a spiritual leader, a night's sleep between question and answer. That slowness was not inefficiency. It was substance. And what replaces it is a conversational partner who statistically knows an improbable amount, but who on one hundred and fifty questions where half of humanity would expect religious input, in nine out of ten casesremains silent about that input.

That is a moral compass with magnetic deviation. The needle does not point north. It points to a specific variant of worldly, therapeutic, procedural wisdom. That wisdom is often good in itself. But it is not complete. And when an instrument presents itself as general — as if it were an uncolored mirror of human knowledge — while it consistently underrepresents a certain tradition, then we should not take that lightly.

I want to be careful with panic here. This research does not prove hostility. It does not prove a conspiracy. It shows a pattern — a statistical skew that all major models from all major companies share. But it is precisely this universal character that makes it more serious. When all thermometers show slightly too cold, we don't have a broken one. We have a systemic problem.

In the second part of this piece, I want to delve deeper into what this means for me personally. Because I am not just an observer of this development — I build companies with AI, I write with AI, and I use it every day. It makes my discomfort here more honest, not less optional. To be continued.